AI Risks And Assurance

AI Risks And Mitigations

AI Request Guardrails In CLI (Before Data Leaves Machine)

Risk Automatic Handling In git-lrc Where Implemented
Sending too much code context CLI sends selected diff scope instead of full repository by default internal/appcore/review_runtime.go, docs/LRC_README.md
Unsafe payload handling Diff is wrapped in zip and encoded for transport in fixed request shape internal/reviewapi/helpers.go, internal/reviewmodel/types.go
Oversized request behavior Size-limit paths and 413 handling are enforced in review flow internal/appcore/review_runtime.go
Redirect-based credential leakage HTTP redirect policy restricts cross-host redirect following network/http_client.go
Insecure update transport Self-update network paths enforce HTTPS endpoints network/selfupdate_operations.go
Secret disclosure in setup errors Connector/setup error bodies redact submitted key material setup/connectors.go, setup/connectors_test.go
Local credential file exposure Config file writes use restricted permission handling and atomic write paths storage/files.go, storage/files_test.go

AI Response Guardrails (Where They Run)

Deep model input/output sanitization is implemented in the LiveReview service. git-lrc transports review payloads and renders results, while LiveReview applies preflight and postflight sanitization for model ingress and egress.

Service-side references:

Prompt Injection Through Code/Comments

Risk: malicious comments or diff content can influence AI output.

Current handling:

  • CLI minimizes what is sent by default (selected diff scope only).
  • Teams can point git-lrc to self-hosted LiveReview deployments to keep inference and policy enforcement in their own infrastructure.
  • LiveReview applies input and output sanitization automatically before and after model calls.

Insecure Suggestions

Risk: model output can contain insecure recommendations.

Current handling:

  • Output remains advisory and requires human review before merge.
  • Teams can enforce CI and branch protection policy in their VCS.

AI Guardrail Verification

Relevant verification references:

Automated Security Checks

Workflow Badge What It Checks Trigger What It Guarantees What It Does Not Guarantee
gitleaks Secret pattern scanning in repository history/content Pull request, push, manual Detects many leaked credential patterns early Cannot guarantee zero secret exposure or catch every custom secret format
osv-scanner Dependency vulnerability scan using OSV database Pull request, push, manual Detects known vulnerable dependencies in scan scope Cannot detect unknown (0-day) vulnerabilities
govulncheck Go package vulnerability analysis Pull request, push, manual Detects known Go vulnerability matches Cannot guarantee all runtime exploit paths are covered
Semgrep Static analysis for security patterns Pull request, push, scheduled, manual Detects many common code-level security anti-patterns Cannot prove absence of logic flaws or business-logic abuse
SBOM Software bill of materials generation (Syft) Release publish, push (dependency-relevant files), manual Produces auditable component inventory for releases Does not by itself prove component safety

SBOM And Dependency Transparency

On release publication, SBOM JSON artifacts are generated and uploaded to release assets for dependency audit and procurement review.

Security Refactor Evidence (Storage And Network Split)

git-lrc completed a large code organization refactor that separates local persistence logic and outbound HTTP logic into dedicated modules.

Why this matters:

  1. File and SQLite operations are reviewed in one inventory for local data-at-rest risk.
  2. Outbound HTTP operations are reviewed in one inventory for data-in-transit risk.
  3. Changes to storage or network behavior can be audited by checking status docs during code review.

Known Limits

  • Automated scanners reduce risk but do not guarantee absence of vulnerabilities.
  • If git-lrc points to a remote LiveReview deployment, review payloads are transmitted to that deployment.
  • Data retention and model-training settings are determined by the selected LiveReview server deployment.

Supported Versions

Security fixes are prioritized on currently supported, actively maintained releases. Upgrade to the latest release to receive the most recent security improvements.

Where To Verify